QV finder: an accurate Quran verse finder system

Abstract

A voice-based search is becoming increasingly important for accessing information across various domains. One of the most challenging areas is Quranic verse search, where precise recitation rules (Tajweed), dialectal variations, and background noise affect accuracy. In this work, we present QV finder, an artificial intelligence (AI)-powered system that utilizes a finetuned whisper-based automatic speech recognition (ASR) model specifically trained on diverse Quranic recitations for the whole Quran. In this paper, we present a robust pipeline for Quranic verse retrieval that bridges the gap between ASR technology and domain-specific linguistic complexity. The model supports both professional and normal reciters, even under noisy conditions. To enhance the localization of verses from partial recitations, we integrated tokenization and advanced string-matching algorithms such as Levenshtein distance and FuzzyWuzzy. For normal reciters, the proposed model achieves a word error rate (WER) of 10.1% and character error rate (CER) of 3.3%, outperforming Google ASR, which exhibits a WER of 27.04%, and a CER of 7.13%. The model also achieves 100% verse retrieval accuracy with a 2.5% false positive rate. Our best fine-tuned model is uploaded here: https://huggingface.co/basharalrfooh/whisper-small-quran.